Source: China Mobile Research Institute
Recently, the knowledge graph-driven communication expert intelligent agent developed by China Mobile Research Institute has been successfully launched at China Mobile Jiangsu Company. This intelligent agent is based on a communication knowledge graph and incorporates a "pattern recognition + large model" hybrid question-and-answer framework, effectively addressing the high barriers of professional knowledge, the difficulty of reusing expert experience, and the issue of isolated cross-domain data in the intelligent transformation of wireless networks. It establishes a new foundation for cognitive decision-making in wireless networks, paving the way for the large-scale application of intelligent agents in existing networks.
The knowledge graph is based on a graph data model, storing multi-source heterogeneous data in a structured form, significantly enhancing the retrieval and reasoning capabilities of complex information, and providing crucial support for research in specialized fields. The China Mobile Research Institute has integrated a "knowledge graph + large model" collaborative framework to independently develop the knowledge graph-driven communication expert intelligent agent: on one hand, it constructs a domain knowledge graph covering 25 categories of communication knowledge, including network architecture, signaling processes, and parameter configurations, with over 100,000 nodes, breaking down cross-domain data silos; on the other hand, it relies on a hybrid framework to balance rule determinism and AI flexibility, establishing a cognitive decision-making foundation for wireless networks. Compared to general large models that require over 500B computing power but still suffer from the absence of specialized knowledge graphs and shallow professional responses, this intelligent agent utilizes a 7B parameter question-and-answer model, achieving a lower hallucination rate than the industry average and a high accuracy of 88.9% under constrained edge computing conditions.
At the level of graph construction, the communication expert intelligent agent relies on entity alignment technology to support cross-language adaptation and dynamic incremental updates, achieving logical associations of domain knowledge; at the model optimization level, it innovatively proposes a hybrid question-and-answer framework collaborative mechanism: the pattern recognition layer uses a rule engine to achieve millisecond-level responses to high-frequency questions; the large model layer interprets complex semantic queries based on a 7B parameter model; the hybrid decision-making mechanism outputs collaboratively through dual engines, enhancing response accuracy. The core innovation of this technology lies in the deep integration of the semantic understanding capability of large models with the structured constraints of knowledge graphs: on one hand, it compensates for the deficiencies of large models in specialized fields through dynamic knowledge sources; on the other hand, it utilizes the logical relationships between knowledge graph entities to assist reasoning, effectively reducing hallucinations from large models.
The knowledge graph-based communication expert intelligent agent has broad application prospects in existing networks: in the area of general technology Q&A, the intelligent agent integrates high-quality domain knowledge to provide traceable signaling process analysis, message parameter queries, and other highly reliable answers, effectively suppressing model hallucinations; in troubleshooting assistance, it generates precise investigation suggestions and more practical solutions based on existing network operation and maintenance case knowledge graphs; in optimizing business experiences, it provides specific performance optimization suggestions through correlation analysis of potential root causes such as terminal measurement status and resource scheduling. With the support of knowledge graphs, this intelligent agent can achieve performance comparable to fully-powered general models with just a 7B parameter model, and in scenarios such as parameter fine-tuning, its accuracy even surpasses that of general models with over ten billion parameters, creating possibilities for the lightweight deployment of intelligent agents on the edge.
As AI RAN technology penetrates all elements of the network, the China Mobile Research Institute is reshaping the architecture of wireless intelligent agents with "knowledge graph + hybrid AI," providing a decision-making engine for existing networks that is "understandable, thorough, and fast in resolution." The research team will continue to promote technological integration, connecting knowledge graphs with existing multi-dimensional intelligent perception assessments and guided fault diagnosis systems to form a closed loop of "perception - analysis - decision-making," accelerating the transformation of technological achievements.
Through the deep integration of communication technology and artificial intelligence, this intelligent agent will possess cognitive reasoning capabilities equivalent to those of communication experts. With knowledge graphs as its core foundation, it will provide a reusable technological basis for the large-scale replication and continuous evolution of network intelligence, injecting strong momentum into the intelligent upgrade of communication networks across the board.返回搜狐，查看更多